HR Agent
An AI platform running in production. The work is under NDA, so this covers the architecture patterns I built and worked with, not the product or its data.
Architecture
- Service-oriented backend
- REST services in front of a relational system of record, each with its own contract and deployment.
- Queue-driven workers
- Background workers consume from queues with claim-check messaging: only an identifier travels, the payload stays in storage.
- Horizontal scale-out
- Replicas coordinate through distributed locks, so the same unit of work is never processed twice.
- Serverless ingestion
- Timer- and queue-triggered functions for ingestion and scheduled jobs.
- Hybrid retrieval
- Vector search combined with keyword search and fused by rank, inside PostgreSQL.
- LLM pipelines
- Batch inference with schema validation and deterministic post-processing, so model output is safe to act on.
- Resilience and cost control
- Retries with back-off, rate limiting, circuit breakers, dry-run modes and spend caps on AI calls.
- Observability
- Distributed tracing and structured logs with OpenTelemetry.
Main processing path
- REST servicescontracts, auth
- queueclaim-check
- workers ×Ndistributed locks
- LLMbatch and real-time
- schema validationsafe to act on
- PostgreSQL + vectorshybrid retrieval
Simplified, with generic component names. Highlighted steps use a model; the rest is software.